attribution patching vs activation patching for a 2-hop factual recall circuit, which is less noisy in practice

trying to localize a 2-hop recall (entity -> attribute -> answer) on an 8b and the two methods disagree on the middle hop. activation patching gives me a clean-ish set of heads around layer 14-19 but its 900+ forward passes and slow. attribution patching (the linear approx from the syrupy paper) is one backward pass but the gradient blows up on exactly the hop i care about, so the second hop lights up half the layers. is that expected because attribution assumes local linearity and the intermediate hop is the least linear part, or am i just under-corrupting the counterfactual. anyone got a repro where attribution and activation actually agree on a multi-hop circuit.